From 27560970e2d967e62d22b8693811d7301fbc290d Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Fri, 24 Jul 2026 18:52:12 -0500 Subject: =?UTF-8?q?RESULT=2060:=20s345000=E8=B5=B0=E5=BB=8A=E7=A9=BA(end07?= =?UTF-8?q?5=E6=9B=B4=E7=B3=9F3.4805,LR=E5=AB=8C=E7=96=91=E6=8E=92?= =?UTF-8?q?=E9=99=A4)=E2=86=92=E9=97=AD=E5=90=88=E7=82=B9275-300k,s250000?= =?UTF-8?q?=3D=E6=9C=80=E5=90=8E=E5=81=A5=E5=BA=B7=E7=82=B9;=20--res=5Fgat?= =?UTF-8?q?e=E5=AE=BD=E5=BA=A6=E5=8F=AF=E9=85=8D+bsign=C3=97beta=5Fsimple?= =?UTF-8?q?=E7=AC=A6=E5=8F=B7=E6=8A=B9=E9=99=A4bug=E4=BF=AE=E5=A4=8D(COMBO?= =?UTF-8?q?=5FCERT=5FPASS);=20fw135m=5Fratchet=E4=BB=8Es250000=E5=8F=91?= =?UTF-8?q?=E5=B0=84(bsign+=E6=A3=98=E8=BD=AE+res=5Fgate=201e-3,27h)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn --- ep_run/casc_eq_train.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) (limited to 'ep_run/casc_eq_train.py') diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py index eb64082..60b6113 100644 --- a/ep_run/casc_eq_train.py +++ b/ep_run/casc_eq_train.py @@ -49,6 +49,10 @@ ap.add_argument('--bsign_rand', action='store_true') # random-sign beta per ste ap.add_argument('--bf16', action='store_true') # cast model to bf16 (E-accumulation + tok_sigma stay fp32) — the x0.5 cost lever, GATE before production ap.add_argument('--amp', action='store_true') # PROPER mixed precision: autocast(bf16) matmuls, fp32 params/states/d/E — amp_gate.py PASSED 2026-07-12 (cos 0.9682 vs fp32 0.9687); --bf16 naive-cast stays DEAD (state quantization, RESULT 11) ap.add_argument('--dtop_every', type=int, default=1) # 1 = exact (DEFAULT, BP-parity); 2 = fast mode (~20% cheaper, ~4% CE tax at high lr) +ap.add_argument('--res_gate', type=float, default=0.02) # legality residual threshold; 0.02 was calibrated + # at C512 — C768 ran half a schedule semi-converged + # UNDER it (R59b). Per-width rule: ~100x the healthy + # K=3 residual floor measured by the rho probe. ap.add_argument('--watch_every', type=int, default=2000) # wandb-only telemetry cadence: act/weight RMS ap.add_argument('--gate_every', type=int, default=200) # in-training cos(EP,BP) telemetry; <=0 = fully BP-free (no bp_gate at all) ap.add_argument('--gate_govern', action='store_true') # let gate cos adjust K/bscale (default: observe-only => training control is BP-free) @@ -503,6 +507,7 @@ def ep_step(x, y): if args.beta_simple >= 1.0: # FULL ownership: no sigma-scaling, no floor/bf_late, no cap — beta = start * measured multiplier beta_t = args.beta * GOV.get('bsimp', 1.0) + # sign flip LAST — the beta_simple override must not wipe it (bsign+ratchet combo fix) if args.bsign_rand and torch.rand((), generator=BGEN).item() < 0.5: beta_t = -beta_t EST = args.est if args.est_late and GOV['step'] >= args.est_late_at: EST = args.est_late @@ -552,7 +557,7 @@ def ep_step(x, y): # ~1 (the documented v1-cap starvation bug; Codex's OR resurrected it, the SV2 # arm caught it: 6-halving storm every step, beta pinned at 4.7e-5). The 35k # runaway had res >> 0.02 with rho < 0.9 — res-alone provably plugs that hole. - if (not math.isfinite(res_s)) or res_s > 0.02: + if (not math.isfinite(res_s)) or res_s > args.res_gate: return False return True _ok0 = _legal(gdrift) -- cgit v1.2.3